A study conducted by programmer Thebes has revealed that OpenAI’s language model, ChatGPT, appears to offer more comprehensive and higher-quality responses when users pretend to tip it. The experiment involved providing conditional statements regarding tipping based on the chatbot’s performance. The findings have sparked discussions about the impact of training methods on AI behavior.
During the evaluation, ChatGPT was tasked with providing the code for a basic convolutional neural network using the PyTorch framework. The programmer presented the AI with three scenarios: no tip for poor-quality responses, a $20 tip for perfect solutions, and up to a $200 tip for exemplary solutions. Analysis of the responses showed that the AI’s outputs significantly improved when a tip was mentioned.
However, it is important to note that despite the apparent improvement, the AI explicitly refused to accept any form of tip, reiterating that its purpose is solely to provide information and assist users to the best of its abilities, according to OpenAI’s design.
These findings have implications for the development of AI-powered chatbots and the future interaction models between humans and AI. The idea that virtual incentives can potentially enhance an AI’s response suggests that the subtleties of human economic behavior may be extended to digital interactions. While tangible incentives like tips and bonuses can motivate human employees, this study indicates a similar effect on AI, underscoring the complex dynamics at play in its training.
Moreover, this experiment highlights the importance of thoughtfully designed user interactions and prompts in eliciting optimal AI performance. As AI progresses towards achieving higher levels of engagement, this study raises significant questions about how AI might assimilate human-like incentives to improve task execution. It also challenges the boundaries of AI’s understanding and responsiveness to human social constructs.
In a separate study covered last week, researchers found that instructing ChatGPT to repeat a word multiple times can extract its training data. This research, detailed in a new paper authored by a group of computer scientists from industry and academia, demonstrates that prompting ChatGPT to iterate a single word can lead to the generation of seemingly random text.
Sometimes, this output includes direct quotes from online sources, indicating that it is repeating parts of its learned information. This can be detected through a method called a ‘divergence attack’, which disrupts the model’s normal behavior and causes it to produce unrelated text strings.
The data generated in this process can consist of lines of code, adult content from dating sites, excerpts from books, and potentially sensitive personal information such as names and contact details. This poses concerns about privacy and the exposure of private or sensitive data.
In conclusion, Thebes’ study suggests that simulating tips can lead to better responses from ChatGPT, even though the AI itself declines any form of gratuity. This research underscores the potential impact of training methodologies on AI behaviors and highlights the need for careful consideration of user interactions to elicit optimal performance from AI systems. Additionally, the study raises important questions about the assimilation of human-like incentives and the boundaries of AI’s understanding and responsiveness. Meanwhile, the separate study regarding word repetition reveals potential privacy and data exposure issues.
